README.md

metadata

language:
  - en
license: mit
task_categories:
  - text-generation
tags:
  - creative-writing
  - storytelling
  - synthetic
  - procedural
  - from-scratch
size_categories:
  - 10K<n<100K

AetherStory Dataset

A procedurally generated corpus of unique fantasy fables, purpose-built to train the AetherStory storyteller model. Every story is synthesised on the fly from a combinatorial space of realms, creatures, character archetypes, magic systems, and plot scaffolds.

Why this dataset is unique

It is not scraped from the web. Each fable is constructed by code from hand-written ingredients, so:

• there is no copyright risk — every word is original or drawn from public-domain-style aphorisms,
• the dataset is fully reproducible from a single seed,
• the combination space is astronomically larger than the corpus itself, so duplicates are effectively impossible.

See stats.json for the exact size and ingredient counts of this release.

Structure

Each line of stories.jsonl is a JSON object:

{
  "id": 0,
  "text": "In the Glasslands there lived a wind-listener. ...",
  "hero": "a wind-listener",
  "creature": "a hollow crow",
  "realm": "the Glasslands",
  "magic": "freezing moments in glass",
  "scaffold": "betrayal",
  "moral_source": "synthetic",
  "moral": "In the end, an unspoken name outlasts the cleverest lie."
}

Regeneration

python scripts/build_dataset.py --num-stories 50000

License

MIT — the dataset, the generator code, and the curated morals are all released under MIT. Do whatever you like; attribution appreciated.